zstd/lib/compress/zstd_lazy.c

1120 lines
51 KiB
C
Raw Normal View History

2017-09-02 01:28:35 +00:00
/*
* Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
* All rights reserved.
*
* This source code is licensed under both the BSD-style license (found in the
* LICENSE file in the root directory of this source tree) and the GPLv2 (found
* in the COPYING file in the root directory of this source tree).
* You may select, at your option, one of the above-listed licenses.
2017-09-02 01:28:35 +00:00
*/
#include "zstd_compress_internal.h"
2017-09-02 01:28:35 +00:00
#include "zstd_lazy.h"
/*-*************************************
* Binary Tree search
***************************************/
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
static void
ZSTD_updateDUBT(ZSTD_matchState_t* ms,
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
const BYTE* ip, const BYTE* iend,
U32 mls)
{
const ZSTD_compressionParameters* const cParams = &ms->cParams;
U32* const hashTable = ms->hashTable;
U32 const hashLog = cParams->hashLog;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32* const bt = ms->chainTable;
U32 const btLog = cParams->chainLog - 1;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 const btMask = (1 << btLog) - 1;
const BYTE* const base = ms->window.base;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 const target = (U32)(ip - base);
U32 idx = ms->nextToUpdate;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
if (idx != target)
DEBUGLOG(7, "ZSTD_updateDUBT, from %u to %u (dictLimit:%u)",
idx, target, ms->window.dictLimit);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
assert(ip + 8 <= iend); /* condition for ZSTD_hashPtr */
(void)iend;
assert(idx >= ms->window.dictLimit); /* condition for valid base+idx */
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
for ( ; idx < target ; idx++) {
size_t const h = ZSTD_hashPtr(base + idx, hashLog, mls); /* assumption : ip + 8 <= iend */
U32 const matchIndex = hashTable[h];
U32* const nextCandidatePtr = bt + 2*(idx&btMask);
U32* const sortMarkPtr = nextCandidatePtr + 1;
DEBUGLOG(8, "ZSTD_updateDUBT: insert %u", idx);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
hashTable[h] = idx; /* Update Hash Table */
*nextCandidatePtr = matchIndex; /* update BT like a chain */
*sortMarkPtr = ZSTD_DUBT_UNSORTED_MARK;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
}
ms->nextToUpdate = target;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
}
/** ZSTD_insertDUBT1() :
* sort one already inserted but unsorted position
* assumption : current >= btlow == (current - btmask)
* doesn't fail */
static void
ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
U32 current, const BYTE* inputEnd,
U32 nbCompares, U32 btLow,
const ZSTD_dictMode_e dictMode)
2017-09-02 01:28:35 +00:00
{
const ZSTD_compressionParameters* const cParams = &ms->cParams;
U32* const bt = ms->chainTable;
U32 const btLog = cParams->chainLog - 1;
U32 const btMask = (1 << btLog) - 1;
2017-09-02 01:28:35 +00:00
size_t commonLengthSmaller=0, commonLengthLarger=0;
const BYTE* const base = ms->window.base;
const BYTE* const dictBase = ms->window.dictBase;
const U32 dictLimit = ms->window.dictLimit;
const BYTE* const ip = (current>=dictLimit) ? base + current : dictBase + current;
const BYTE* const iend = (current>=dictLimit) ? inputEnd : dictBase + dictLimit;
2017-09-02 01:28:35 +00:00
const BYTE* const dictEnd = dictBase + dictLimit;
const BYTE* const prefixStart = base + dictLimit;
const BYTE* match;
U32* smallerPtr = bt + 2*(current&btMask);
U32* largerPtr = smallerPtr + 1;
U32 matchIndex = *smallerPtr; /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */
2017-09-02 01:28:35 +00:00
U32 dummy32; /* to be nullified at the end */
U32 const windowValid = ms->window.lowLimit;
U32 const maxDistance = 1U << cParams->windowLog;
U32 const windowLow = (current - windowValid > maxDistance) ? current - maxDistance : windowValid;
2017-09-02 01:28:35 +00:00
DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
current, dictLimit, windowLow);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
assert(current >= btLow);
assert(ip < iend); /* condition for ZSTD_count */
2017-09-02 01:28:35 +00:00
while (nbCompares-- && (matchIndex > windowLow)) {
U32* const nextPtr = bt + 2*(matchIndex & btMask);
size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */
Fixed Btree update ZSTD_updateTree() expected to be followed by a Bt match finder, which would update zc->nextToUpdate. With the new optimal match finder, it's not necessarily the case : a match might be found during repcode or hash3, and stops there because it reaches sufficient_len, without even entering the binary tree. Previous policy was to nonetheless update zc->nextToUpdate, but the current position would not be inserted, creating "holes" in the btree, aka positions that will no longer be searched. Now, when current position is not inserted, zc->nextToUpdate is not update, expecting ZSTD_updateTree() to fill the tree later on. Solution selected is that ZSTD_updateTree() takes care of properly setting zc->nextToUpdate, so that it no longer depends on a future function to do this job. It took time to get there, as the issue started with a memory sanitizer error. The pb would have been easier to spot with a proper `assert()`. So this patch add a few of them. Additionnally, I discovered that `make test` does not enable `assert()` during CLI tests. This patch enables them. Unfortunately, these `assert()` triggered other (unrelated) bugs during CLI tests, mostly within zstdmt. So this patch also fixes them. - Changed packed structure for gcc memory access : memory sanitizer would complain that a read "might" reach out-of-bound position on the ground that the `union` is larger than the type accessed. Now, to avoid this issue, each type is independent. - ZSTD_CCtxParams_setParameter() : @return provides the value of parameter, clamped/fixed appropriately. - ZSTDMT : changed constant name to ZSTDMT_JOBSIZE_MIN - ZSTDMT : multithreading is automatically disabled when srcSize <= ZSTDMT_JOBSIZE_MIN, since only one thread will be used in this case (saves memory and runtime). - ZSTDMT : nbThreads is automatically clamped on setting the value.
2017-11-16 20:18:56 +00:00
assert(matchIndex < current);
/* note : all candidates are now supposed sorted,
* but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK
* when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */
2017-09-02 01:28:35 +00:00
if ( (dictMode != ZSTD_extDict)
|| (matchIndex+matchLength >= dictLimit) /* both in current segment*/
|| (current < dictLimit) /* both in extDict */) {
const BYTE* const mBase = ( (dictMode != ZSTD_extDict)
|| (matchIndex+matchLength >= dictLimit)) ?
base : dictBase;
assert( (matchIndex+matchLength >= dictLimit) /* might be wrong if extDict is incorrectly set to 0 */
|| (current < dictLimit) );
match = mBase + matchIndex;
2017-11-19 22:40:21 +00:00
matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
2017-09-02 01:28:35 +00:00
} else {
match = dictBase + matchIndex;
matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
if (matchIndex+matchLength >= dictLimit)
match = base + matchIndex; /* preparation for next read of match[matchLength] */
2017-09-02 01:28:35 +00:00
}
DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
current, matchIndex, (U32)matchLength);
if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */
2017-09-17 06:40:14 +00:00
break; /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
}
2017-09-02 01:28:35 +00:00
2017-09-17 06:40:14 +00:00
if (match[matchLength] < ip[matchLength]) { /* necessarily within buffer */
2017-11-15 21:44:24 +00:00
/* match is smaller than current */
2017-09-02 01:28:35 +00:00
*smallerPtr = matchIndex; /* update smaller idx */
commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */
2017-09-17 06:40:14 +00:00
if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop searching */
DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is smaller : next => %u",
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
matchIndex, btLow, nextPtr[1]);
2017-11-15 21:44:24 +00:00
smallerPtr = nextPtr+1; /* new "candidate" => larger than match, which was smaller than target */
matchIndex = nextPtr[1]; /* new matchIndex, larger than previous and closer to current */
2017-09-02 01:28:35 +00:00
} else {
/* match is larger than current */
*largerPtr = matchIndex;
commonLengthLarger = matchLength;
2017-09-17 06:40:14 +00:00
if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop searching */
DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is larger => %u",
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
matchIndex, btLow, nextPtr[0]);
2017-09-02 01:28:35 +00:00
largerPtr = nextPtr;
matchIndex = nextPtr[0];
} }
*smallerPtr = *largerPtr = 0;
Fixed Btree update ZSTD_updateTree() expected to be followed by a Bt match finder, which would update zc->nextToUpdate. With the new optimal match finder, it's not necessarily the case : a match might be found during repcode or hash3, and stops there because it reaches sufficient_len, without even entering the binary tree. Previous policy was to nonetheless update zc->nextToUpdate, but the current position would not be inserted, creating "holes" in the btree, aka positions that will no longer be searched. Now, when current position is not inserted, zc->nextToUpdate is not update, expecting ZSTD_updateTree() to fill the tree later on. Solution selected is that ZSTD_updateTree() takes care of properly setting zc->nextToUpdate, so that it no longer depends on a future function to do this job. It took time to get there, as the issue started with a memory sanitizer error. The pb would have been easier to spot with a proper `assert()`. So this patch add a few of them. Additionnally, I discovered that `make test` does not enable `assert()` during CLI tests. This patch enables them. Unfortunately, these `assert()` triggered other (unrelated) bugs during CLI tests, mostly within zstdmt. So this patch also fixes them. - Changed packed structure for gcc memory access : memory sanitizer would complain that a read "might" reach out-of-bound position on the ground that the `union` is larger than the type accessed. Now, to avoid this issue, each type is independent. - ZSTD_CCtxParams_setParameter() : @return provides the value of parameter, clamped/fixed appropriately. - ZSTDMT : changed constant name to ZSTDMT_JOBSIZE_MIN - ZSTDMT : multithreading is automatically disabled when srcSize <= ZSTDMT_JOBSIZE_MIN, since only one thread will be used in this case (saves memory and runtime). - ZSTDMT : nbThreads is automatically clamped on setting the value.
2017-11-16 20:18:56 +00:00
}
2017-09-02 01:28:35 +00:00
static size_t
ZSTD_DUBT_findBetterDictMatch (
ZSTD_matchState_t* ms,
2018-06-12 22:38:10 +00:00
const BYTE* const ip, const BYTE* const iend,
size_t* offsetPtr,
size_t bestLength,
2018-06-12 22:38:10 +00:00
U32 nbCompares,
U32 const mls,
const ZSTD_dictMode_e dictMode)
{
2018-06-12 22:38:10 +00:00
const ZSTD_matchState_t * const dms = ms->dictMatchState;
const ZSTD_compressionParameters* const dmsCParams = &dms->cParams;
2018-06-12 22:38:10 +00:00
const U32 * const dictHashTable = dms->hashTable;
U32 const hashLog = dmsCParams->hashLog;
2018-06-12 22:38:10 +00:00
size_t const h = ZSTD_hashPtr(ip, hashLog, mls);
U32 dictMatchIndex = dictHashTable[h];
const BYTE* const base = ms->window.base;
const BYTE* const prefixStart = base + ms->window.dictLimit;
U32 const current = (U32)(ip-base);
const BYTE* const dictBase = dms->window.base;
const BYTE* const dictEnd = dms->window.nextSrc;
U32 const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base);
U32 const dictLowLimit = dms->window.lowLimit;
U32 const dictIndexDelta = ms->window.lowLimit - dictHighLimit;
U32* const dictBt = dms->chainTable;
U32 const btLog = dmsCParams->chainLog - 1;
2018-06-12 22:38:10 +00:00
U32 const btMask = (1 << btLog) - 1;
2018-06-21 19:24:08 +00:00
U32 const btLow = (btMask >= dictHighLimit - dictLowLimit) ? dictLowLimit : dictHighLimit - btMask;
2018-06-12 22:38:10 +00:00
size_t commonLengthSmaller=0, commonLengthLarger=0;
2018-06-12 22:38:10 +00:00
(void)dictMode;
assert(dictMode == ZSTD_dictMatchState);
while (nbCompares-- && (dictMatchIndex > dictLowLimit)) {
U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */
const BYTE* match = dictBase + dictMatchIndex;
matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
if (dictMatchIndex+matchLength >= dictHighLimit)
match = base + dictMatchIndex + dictIndexDelta; /* to prepare for next usage of match[matchLength] */
if (matchLength > bestLength) {
U32 matchIndex = dictMatchIndex + dictIndexDelta;
if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
2018-10-08 22:50:02 +00:00
DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
2018-06-12 22:38:10 +00:00
current, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, ZSTD_REP_MOVE + current - matchIndex, dictMatchIndex, matchIndex);
bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex;
}
if (ip+matchLength == iend) { /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
2018-06-12 22:38:10 +00:00
break; /* drop, to guarantee consistency (miss a little bit of compression) */
}
}
if (match[matchLength] < ip[matchLength]) {
if (dictMatchIndex <= btLow) { break; } /* beyond tree size, stop the search */
commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */
dictMatchIndex = nextPtr[1]; /* new matchIndex larger than previous (closer to current) */
} else {
/* match is larger than current */
if (dictMatchIndex <= btLow) { break; } /* beyond tree size, stop the search */
commonLengthLarger = matchLength;
dictMatchIndex = nextPtr[0];
}
}
if (bestLength >= MINMATCH) {
U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
2018-10-08 22:50:02 +00:00
DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
2018-06-12 22:38:10 +00:00
current, (U32)bestLength, (U32)*offsetPtr, mIndex);
}
return bestLength;
}
static size_t
ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
const BYTE* const ip, const BYTE* const iend,
size_t* offsetPtr,
U32 const mls,
const ZSTD_dictMode_e dictMode)
2017-09-02 01:28:35 +00:00
{
const ZSTD_compressionParameters* const cParams = &ms->cParams;
U32* const hashTable = ms->hashTable;
U32 const hashLog = cParams->hashLog;
2017-09-02 01:28:35 +00:00
size_t const h = ZSTD_hashPtr(ip, hashLog, mls);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 matchIndex = hashTable[h];
const BYTE* const base = ms->window.base;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 const current = (U32)(ip-base);
U32 const maxDistance = 1U << cParams->windowLog;
2019-08-02 12:26:26 +00:00
U32 const lowestValid = ms->window.lowLimit;
U32 const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
U32 const isDictionary = (ms->loadedDictEnd != 0);
U32 const windowLow = isDictionary ? lowestValid : withinWindow;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32* const bt = ms->chainTable;
U32 const btLog = cParams->chainLog - 1;
2017-09-02 01:28:35 +00:00
U32 const btMask = (1 << btLog) - 1;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 const btLow = (btMask >= current) ? 0 : current - btMask;
U32 const unsortLimit = MAX(btLow, windowLow);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32* nextCandidate = bt + 2*(matchIndex&btMask);
U32* unsortedMark = bt + 2*(matchIndex&btMask) + 1;
U32 nbCompares = 1U << cParams->searchLog;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 nbCandidates = nbCompares;
U32 previousCandidate = 0;
2017-09-02 01:28:35 +00:00
DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", current);
2017-09-17 06:40:14 +00:00
assert(ip <= iend-8); /* required for h calculation */
2017-09-02 01:28:35 +00:00
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
/* reach end of unsorted candidates list */
while ( (matchIndex > unsortLimit)
&& (*unsortedMark == ZSTD_DUBT_UNSORTED_MARK)
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
&& (nbCandidates > 1) ) {
DEBUGLOG(8, "ZSTD_DUBT_findBestMatch: candidate %u is unsorted",
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
matchIndex);
*unsortedMark = previousCandidate; /* the unsortedMark becomes a reversed chain, to move up back to original position */
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
previousCandidate = matchIndex;
matchIndex = *nextCandidate;
nextCandidate = bt + 2*(matchIndex&btMask);
unsortedMark = bt + 2*(matchIndex&btMask) + 1;
nbCandidates --;
}
2017-09-02 01:28:35 +00:00
/* nullify last candidate if it's still unsorted
* simplification, detrimental to compression ratio, beneficial for speed */
if ( (matchIndex > unsortLimit)
&& (*unsortedMark==ZSTD_DUBT_UNSORTED_MARK) ) {
DEBUGLOG(7, "ZSTD_DUBT_findBestMatch: nullify last unsorted candidate %u",
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
matchIndex);
*nextCandidate = *unsortedMark = 0;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
}
2017-09-02 01:28:35 +00:00
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
/* batch sort stacked candidates */
matchIndex = previousCandidate;
while (matchIndex) { /* will end on matchIndex == 0 */
U32* const nextCandidateIdxPtr = bt + 2*(matchIndex&btMask) + 1;
U32 const nextCandidateIdx = *nextCandidateIdxPtr;
ZSTD_insertDUBT1(ms, matchIndex, iend,
nbCandidates, unsortLimit, dictMode);
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
matchIndex = nextCandidateIdx;
nbCandidates++;
}
2017-09-02 01:28:35 +00:00
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
/* find longest match */
{ size_t commonLengthSmaller = 0, commonLengthLarger = 0;
const BYTE* const dictBase = ms->window.dictBase;
const U32 dictLimit = ms->window.dictLimit;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
const BYTE* const dictEnd = dictBase + dictLimit;
const BYTE* const prefixStart = base + dictLimit;
U32* smallerPtr = bt + 2*(current&btMask);
U32* largerPtr = bt + 2*(current&btMask) + 1;
U32 matchEndIdx = current + 8 + 1;
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
U32 dummy32; /* to be nullified at the end */
size_t bestLength = 0;
matchIndex = hashTable[h];
hashTable[h] = current; /* Update Hash Table */
while (nbCompares-- && (matchIndex > windowLow)) {
U32* const nextPtr = bt + 2*(matchIndex & btMask);
size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */
const BYTE* match;
if ((dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit)) {
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
match = base + matchIndex;
matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
} else {
match = dictBase + matchIndex;
matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
if (matchIndex+matchLength >= dictLimit)
match = base + matchIndex; /* to prepare for next usage of match[matchLength] */
}
2017-09-02 01:28:35 +00:00
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
if (matchLength > bestLength) {
if (matchLength > matchEndIdx - matchIndex)
matchEndIdx = matchIndex + (U32)matchLength;
if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex;
if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */
if (dictMode == ZSTD_dictMatchState) {
nbCompares = 0; /* in addition to avoiding checking any
* further in this loop, make sure we
* skip checking in the dictionary. */
}
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
break; /* drop, to guarantee consistency (miss a little bit of compression) */
}
}
2017-09-02 01:28:35 +00:00
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
if (match[matchLength] < ip[matchLength]) {
/* match is smaller than current */
*smallerPtr = matchIndex; /* update smaller idx */
commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */
if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop the search */
smallerPtr = nextPtr+1; /* new "smaller" => larger of match */
matchIndex = nextPtr[1]; /* new matchIndex larger than previous (closer to current) */
} else {
/* match is larger than current */
*largerPtr = matchIndex;
commonLengthLarger = matchLength;
if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop the search */
largerPtr = nextPtr;
matchIndex = nextPtr[0];
} }
*smallerPtr = *largerPtr = 0;
2018-06-12 22:38:10 +00:00
if (dictMode == ZSTD_dictMatchState && nbCompares) {
bestLength = ZSTD_DUBT_findBetterDictMatch(
ms, ip, iend,
offsetPtr, bestLength, nbCompares,
mls, dictMode);
2018-06-12 22:38:10 +00:00
}
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
assert(matchEndIdx > current+8); /* ensure nextToUpdate is increased */
ms->nextToUpdate = matchEndIdx - 8; /* skip repetitive patterns */
if (bestLength >= MINMATCH) {
U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
current, (U32)bestLength, (U32)*offsetPtr, mIndex);
}
first implementation of delayed update for btlazy2 This is a pretty nice speed win. The new strategy consists in stacking new candidates as if it was a hash chain. Then, only if there is a need to actually consult the chain, they are batch-updated, before starting the match search itself. This is supposed to be beneficial when skipping positions, which happens a lot when using lazy strategy. The baseline performance for btlazy2 on my laptop is : 15#calgary.tar : 3265536 -> 955985 (3.416), 7.06 MB/s , 618.0 MB/s 15#enwik7 : 10000000 -> 3067341 (3.260), 4.65 MB/s , 521.2 MB/s 15#silesia.tar : 211984896 -> 58095131 (3.649), 6.20 MB/s , 682.4 MB/s (only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt) After this patch, and keeping all parameters identical, speed is increased by a pretty good margin (+30-50%), but compression ratio suffers a bit : 15#calgary.tar : 3265536 -> 958060 (3.408), 9.12 MB/s , 621.1 MB/s 15#enwik7 : 10000000 -> 3078318 (3.249), 6.37 MB/s , 525.1 MB/s 15#silesia.tar : 211984896 -> 58444111 (3.627), 9.89 MB/s , 680.4 MB/s That's because I kept `1<<searchLog` as a maximum number of candidates to update. But for a hash chain, this represents the total number of candidates in the chain, while for the binary, it represents the maximum depth of searches. Keep in mind that a lot of candidates won't even be visited in the btree, since they are filtered out by the binary sort. As a consequence, in the new implementation, the effective depth of the binary tree is substantially shorter. To compensate, it's enough to increase `searchLog` value. Here is the result after adding just +1 to searchLog (level 15 setting in this patch): 15#calgary.tar : 3265536 -> 956311 (3.415), 8.32 MB/s , 611.4 MB/s 15#enwik7 : 10000000 -> 3067655 (3.260), 5.43 MB/s , 535.5 MB/s 15#silesia.tar : 211984896 -> 58113144 (3.648), 8.35 MB/s , 679.3 MB/s aka, almost the same compression ratio as before, but with a noticeable speed increase (+20-30%). This modification makes btlazy2 more competitive. A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
2017-12-28 15:58:57 +00:00
return bestLength;
}
2017-09-02 01:28:35 +00:00
}
/** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
FORCE_INLINE_TEMPLATE size_t
ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
const BYTE* const ip, const BYTE* const iLimit,
size_t* offsetPtr,
const U32 mls /* template */,
const ZSTD_dictMode_e dictMode)
2017-09-02 01:28:35 +00:00
{
DEBUGLOG(7, "ZSTD_BtFindBestMatch");
if (ip < ms->window.base + ms->nextToUpdate) return 0; /* skipped area */
ZSTD_updateDUBT(ms, ip, iLimit, mls);
return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
2017-09-02 01:28:35 +00:00
}
static size_t
ZSTD_BtFindBestMatch_selectMLS ( ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
{
switch(ms->cParams.minMatch)
{
default : /* includes case 3 */
case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
case 7 :
case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
}
}
static size_t ZSTD_BtFindBestMatch_dictMatchState_selectMLS (
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
{
switch(ms->cParams.minMatch)
{
default : /* includes case 3 */
case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
case 7 :
case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
}
}
static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
2017-09-02 01:28:35 +00:00
{
switch(ms->cParams.minMatch)
2017-09-02 01:28:35 +00:00
{
default : /* includes case 3 */
case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
2017-09-02 01:28:35 +00:00
case 7 :
case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
2017-09-02 01:28:35 +00:00
}
}
/* *********************************
* Hash Chain
***********************************/
#define NEXT_IN_CHAIN(d, mask) chainTable[(d) & (mask)]
2017-09-02 01:28:35 +00:00
/* Update chains up to ip (excluded)
Assumption : always within prefix (i.e. not within extDict) */
static U32 ZSTD_insertAndFindFirstIndex_internal(
ZSTD_matchState_t* ms,
const ZSTD_compressionParameters* const cParams,
const BYTE* ip, U32 const mls)
2017-09-02 01:28:35 +00:00
{
U32* const hashTable = ms->hashTable;
const U32 hashLog = cParams->hashLog;
U32* const chainTable = ms->chainTable;
const U32 chainMask = (1 << cParams->chainLog) - 1;
const BYTE* const base = ms->window.base;
2017-09-02 01:28:35 +00:00
const U32 target = (U32)(ip - base);
U32 idx = ms->nextToUpdate;
2017-09-02 01:28:35 +00:00
while(idx < target) { /* catch up */
size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
hashTable[h] = idx;
idx++;
}
ms->nextToUpdate = target;
2017-09-02 01:28:35 +00:00
return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
}
U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
const ZSTD_compressionParameters* const cParams = &ms->cParams;
return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
}
2017-09-02 01:28:35 +00:00
/* inlining is important to hardwire a hot branch (template emulation) */
FORCE_INLINE_TEMPLATE
size_t ZSTD_HcFindBestMatch_generic (
ZSTD_matchState_t* ms,
2017-09-02 01:28:35 +00:00
const BYTE* const ip, const BYTE* const iLimit,
size_t* offsetPtr,
const U32 mls, const ZSTD_dictMode_e dictMode)
2017-09-02 01:28:35 +00:00
{
const ZSTD_compressionParameters* const cParams = &ms->cParams;
U32* const chainTable = ms->chainTable;
const U32 chainSize = (1 << cParams->chainLog);
2017-09-02 01:28:35 +00:00
const U32 chainMask = chainSize-1;
const BYTE* const base = ms->window.base;
const BYTE* const dictBase = ms->window.dictBase;
const U32 dictLimit = ms->window.dictLimit;
2017-09-02 01:28:35 +00:00
const BYTE* const prefixStart = base + dictLimit;
const BYTE* const dictEnd = dictBase + dictLimit;
const U32 current = (U32)(ip-base);
const U32 maxDistance = 1U << cParams->windowLog;
const U32 lowestValid = ms->window.lowLimit;
const U32 withinMaxDistance = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
const U32 isDictionary = (ms->loadedDictEnd != 0);
const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
2017-09-02 01:28:35 +00:00
const U32 minChain = current > chainSize ? current - chainSize : 0;
U32 nbAttempts = 1U << cParams->searchLog;
2017-09-02 01:28:35 +00:00
size_t ml=4-1;
/* HC4 match finder */
U32 matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
2017-09-02 01:28:35 +00:00
for ( ; (matchIndex>lowLimit) & (nbAttempts>0) ; nbAttempts--) {
size_t currentMl=0;
if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
2017-11-19 22:40:21 +00:00
const BYTE* const match = base + matchIndex;
assert(matchIndex >= dictLimit); /* ensures this is true if dictMode != ZSTD_extDict */
2017-09-02 01:28:35 +00:00
if (match[ml] == ip[ml]) /* potentially better */
currentMl = ZSTD_count(ip, match, iLimit);
} else {
2017-11-19 22:40:21 +00:00
const BYTE* const match = dictBase + matchIndex;
assert(match+4 <= dictEnd);
2017-09-02 01:28:35 +00:00
if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */
currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
}
/* save best solution */
if (currentMl > ml) {
ml = currentMl;
*offsetPtr = current - matchIndex + ZSTD_REP_MOVE;
if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
}
if (matchIndex <= minChain) break;
matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
}
if (dictMode == ZSTD_dictMatchState) {
const ZSTD_matchState_t* const dms = ms->dictMatchState;
const U32* const dmsChainTable = dms->chainTable;
const U32 dmsChainSize = (1 << dms->cParams.chainLog);
const U32 dmsChainMask = dmsChainSize - 1;
const U32 dmsLowestIndex = dms->window.dictLimit;
const BYTE* const dmsBase = dms->window.base;
const BYTE* const dmsEnd = dms->window.nextSrc;
const U32 dmsSize = (U32)(dmsEnd - dmsBase);
const U32 dmsIndexDelta = dictLimit - dmsSize;
const U32 dmsMinChain = dmsSize > dmsChainSize ? dmsSize - dmsChainSize : 0;
matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];
for ( ; (matchIndex>dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
size_t currentMl=0;
const BYTE* const match = dmsBase + matchIndex;
assert(match+4 <= dmsEnd);
if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */
currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
/* save best solution */
if (currentMl > ml) {
ml = currentMl;
*offsetPtr = current - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
}
if (matchIndex <= dmsMinChain) break;
matchIndex = dmsChainTable[matchIndex & dmsChainMask];
}
}
2017-09-02 01:28:35 +00:00
return ml;
}
FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS (
ZSTD_matchState_t* ms,
2017-09-02 01:28:35 +00:00
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
{
switch(ms->cParams.minMatch)
{
default : /* includes case 3 */
case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
case 7 :
case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
}
}
2018-06-07 20:55:13 +00:00
static size_t ZSTD_HcFindBestMatch_dictMatchState_selectMLS (
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
{
switch(ms->cParams.minMatch)
{
default : /* includes case 3 */
case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
case 7 :
case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
}
}
FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* const iLimit,
size_t* offsetPtr)
2017-09-02 01:28:35 +00:00
{
switch(ms->cParams.minMatch)
2017-09-02 01:28:35 +00:00
{
default : /* includes case 3 */
case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
2017-09-02 01:28:35 +00:00
case 7 :
case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
2017-09-02 01:28:35 +00:00
}
}
/* *******************************
* Common parser - lazy strategy
*********************************/
typedef enum { search_hashChain, search_binaryTree } searchMethod_e;
FORCE_INLINE_TEMPLATE size_t
ZSTD_compressBlock_lazy_generic(
ZSTD_matchState_t* ms, seqStore_t* seqStore,
U32 rep[ZSTD_REP_NUM],
const void* src, size_t srcSize,
const searchMethod_e searchMethod, const U32 depth,
ZSTD_dictMode_e const dictMode)
2017-09-02 01:28:35 +00:00
{
const BYTE* const istart = (const BYTE*)src;
const BYTE* ip = istart;
const BYTE* anchor = istart;
const BYTE* const iend = istart + srcSize;
const BYTE* const ilimit = iend - 8;
2018-05-23 19:06:24 +00:00
const BYTE* const base = ms->window.base;
const U32 prefixLowestIndex = ms->window.dictLimit;
const BYTE* const prefixLowest = base + prefixLowestIndex;
2017-09-02 01:28:35 +00:00
typedef size_t (*searchMax_f)(
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
searchMax_f const searchMax = dictMode == ZSTD_dictMatchState ?
(searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_dictMatchState_selectMLS
: ZSTD_HcFindBestMatch_dictMatchState_selectMLS) :
(searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_selectMLS
: ZSTD_HcFindBestMatch_selectMLS);
U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
2017-09-02 01:28:35 +00:00
const ZSTD_matchState_t* const dms = ms->dictMatchState;
const U32 dictLowestIndex = dictMode == ZSTD_dictMatchState ?
dms->window.dictLimit : 0;
const BYTE* const dictBase = dictMode == ZSTD_dictMatchState ?
dms->window.base : NULL;
const BYTE* const dictLowest = dictMode == ZSTD_dictMatchState ?
dictBase + dictLowestIndex : NULL;
const BYTE* const dictEnd = dictMode == ZSTD_dictMatchState ?
dms->window.nextSrc : NULL;
2018-05-23 19:06:24 +00:00
const U32 dictIndexDelta = dictMode == ZSTD_dictMatchState ?
prefixLowestIndex - (U32)(dictEnd - dictBase) :
0;
const U32 dictAndPrefixLength = (U32)(ip - prefixLowest + dictEnd - dictLowest);
2017-09-02 01:28:35 +00:00
/* init */
ip += (dictAndPrefixLength == 0);
if (dictMode == ZSTD_noDict) {
U32 const maxRep = (U32)(ip - prefixLowest);
2017-09-02 01:28:35 +00:00
if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
}
if (dictMode == ZSTD_dictMatchState) {
/* dictMatchState repCode checks don't currently handle repCode == 0
* disabling. */
assert(offset_1 <= dictAndPrefixLength);
assert(offset_2 <= dictAndPrefixLength);
}
2017-09-02 01:28:35 +00:00
/* Match Loop */
while (ip < ilimit) {
size_t matchLength=0;
size_t offset=0;
const BYTE* start=ip+1;
/* check repCode */
2018-06-06 23:54:13 +00:00
if (dictMode == ZSTD_dictMatchState) {
const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
const BYTE* repMatch = (dictMode == ZSTD_dictMatchState
&& repIndex < prefixLowestIndex) ?
dictBase + (repIndex - dictIndexDelta) :
base + repIndex;
if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
&& (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
if (depth==0) goto _storeSequence;
}
}
if ( dictMode == ZSTD_noDict
&& ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
2018-05-23 19:06:24 +00:00
matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
2017-09-02 01:28:35 +00:00
if (depth==0) goto _storeSequence;
}
/* first search (depth 0) */
{ size_t offsetFound = 999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
2017-09-02 01:28:35 +00:00
if (ml2 > matchLength)
matchLength = ml2, start = ip, offset=offsetFound;
}
if (matchLength < 4) {
ip += ((ip-anchor) >> kSearchStrength) + 1; /* jump faster over incompressible sections */
2017-09-02 01:28:35 +00:00
continue;
}
/* let's try to find a better solution */
if (depth>=1)
while (ip<ilimit) {
ip ++;
2018-06-06 23:54:13 +00:00
if ( (dictMode == ZSTD_noDict)
&& (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
2017-09-02 01:28:35 +00:00
size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
int const gain2 = (int)(mlRep * 3);
int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
if ((mlRep >= 4) && (gain2 > gain1))
matchLength = mlRep, offset = 0, start = ip;
}
2018-06-06 23:54:13 +00:00
if (dictMode == ZSTD_dictMatchState) {
const U32 repIndex = (U32)(ip - base) - offset_1;
const BYTE* repMatch = repIndex < prefixLowestIndex ?
dictBase + (repIndex - dictIndexDelta) :
base + repIndex;
if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
&& (MEM_read32(repMatch) == MEM_read32(ip)) ) {
const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
int const gain2 = (int)(mlRep * 3);
int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
if ((mlRep >= 4) && (gain2 > gain1))
matchLength = mlRep, offset = 0, start = ip;
}
}
{ size_t offset2=999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offset2);
2017-09-02 01:28:35 +00:00
int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
if ((ml2 >= 4) && (gain2 > gain1)) {
matchLength = ml2, offset = offset2, start = ip;
continue; /* search a better one */
} }
/* let's find an even better one */
if ((depth==2) && (ip<ilimit)) {
ip ++;
2018-06-06 23:54:13 +00:00
if ( (dictMode == ZSTD_noDict)
&& (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
int const gain2 = (int)(mlRep * 4);
2017-09-02 01:28:35 +00:00
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
2018-06-06 23:54:13 +00:00
if ((mlRep >= 4) && (gain2 > gain1))
matchLength = mlRep, offset = 0, start = ip;
}
if (dictMode == ZSTD_dictMatchState) {
const U32 repIndex = (U32)(ip - base) - offset_1;
const BYTE* repMatch = repIndex < prefixLowestIndex ?
dictBase + (repIndex - dictIndexDelta) :
base + repIndex;
if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
&& (MEM_read32(repMatch) == MEM_read32(ip)) ) {
const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
int const gain2 = (int)(mlRep * 4);
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
if ((mlRep >= 4) && (gain2 > gain1))
matchLength = mlRep, offset = 0, start = ip;
}
2017-09-02 01:28:35 +00:00
}
{ size_t offset2=999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offset2);
2017-09-02 01:28:35 +00:00
int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
if ((ml2 >= 4) && (gain2 > gain1)) {
matchLength = ml2, offset = offset2, start = ip;
continue;
} } }
break; /* nothing found : store previous solution */
}
/* NOTE:
* start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
* (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
* overflows the pointer, which is undefined behavior.
*/
/* catch up */
if (offset) {
2018-06-08 19:06:47 +00:00
if (dictMode == ZSTD_noDict) {
while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest))
&& (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) ) /* only search for offset within prefix */
{ start--; matchLength++; }
}
if (dictMode == ZSTD_dictMatchState) {
U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */
}
2017-09-02 01:28:35 +00:00
offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
}
/* store sequence */
_storeSequence:
{ size_t const litLength = start - anchor;
ZSTD_storeSeq(seqStore, litLength, anchor, (U32)offset, matchLength-MINMATCH);
2017-09-02 01:28:35 +00:00
anchor = ip = start + matchLength;
}
/* check immediate repcode */
2018-05-23 19:49:43 +00:00
if (dictMode == ZSTD_dictMatchState) {
while (ip <= ilimit) {
U32 const current2 = (U32)(ip-base);
2018-06-06 23:54:13 +00:00
U32 const repIndex = current2 - offset_2;
const BYTE* repMatch = dictMode == ZSTD_dictMatchState
&& repIndex < prefixLowestIndex ?
dictBase - dictIndexDelta + repIndex :
base + repIndex;
if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
&& (MEM_read32(repMatch) == MEM_read32(ip)) ) {
const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
2018-06-09 21:19:04 +00:00
offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset_2 <=> offset_1 */
2018-05-23 19:49:43 +00:00
ZSTD_storeSeq(seqStore, 0, anchor, 0, matchLength-MINMATCH);
ip += matchLength;
anchor = ip;
continue;
}
break;
}
}
if (dictMode == ZSTD_noDict) {
while ( ((ip <= ilimit) & (offset_2>0))
&& (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
/* store sequence */
matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
ZSTD_storeSeq(seqStore, 0, anchor, 0, matchLength-MINMATCH);
ip += matchLength;
anchor = ip;
continue; /* faster when present ... (?) */
} } }
2017-09-02 01:28:35 +00:00
/* Save reps for next block */
rep[0] = offset_1 ? offset_1 : savedOffset;
rep[1] = offset_2 ? offset_2 : savedOffset;
2017-09-02 01:28:35 +00:00
/* Return the last literals size */
2019-08-02 12:42:53 +00:00
return (size_t)(iend - anchor);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_btlazy2(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_lazy2(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_lazy(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_greedy(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_btlazy2_dictMatchState(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
}
size_t ZSTD_compressBlock_lazy2_dictMatchState(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
}
size_t ZSTD_compressBlock_lazy_dictMatchState(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
}
size_t ZSTD_compressBlock_greedy_dictMatchState(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
{
return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
}
2017-09-02 01:28:35 +00:00
FORCE_INLINE_TEMPLATE
size_t ZSTD_compressBlock_lazy_extDict_generic(
ZSTD_matchState_t* ms, seqStore_t* seqStore,
U32 rep[ZSTD_REP_NUM],
const void* src, size_t srcSize,
const searchMethod_e searchMethod, const U32 depth)
2017-09-02 01:28:35 +00:00
{
const BYTE* const istart = (const BYTE*)src;
const BYTE* ip = istart;
const BYTE* anchor = istart;
const BYTE* const iend = istart + srcSize;
const BYTE* const ilimit = iend - 8;
const BYTE* const base = ms->window.base;
const U32 dictLimit = ms->window.dictLimit;
const U32 lowestIndex = ms->window.lowLimit;
2017-09-02 01:28:35 +00:00
const BYTE* const prefixStart = base + dictLimit;
const BYTE* const dictBase = ms->window.dictBase;
2017-09-02 01:28:35 +00:00
const BYTE* const dictEnd = dictBase + dictLimit;
const BYTE* const dictStart = dictBase + lowestIndex;
2017-09-02 01:28:35 +00:00
typedef size_t (*searchMax_f)(
ZSTD_matchState_t* ms,
const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS;
2017-09-02 01:28:35 +00:00
U32 offset_1 = rep[0], offset_2 = rep[1];
2017-09-02 01:28:35 +00:00
/* init */
ip += (ip == prefixStart);
/* Match Loop */
while (ip < ilimit) {
size_t matchLength=0;
size_t offset=0;
const BYTE* start=ip+1;
U32 current = (U32)(ip-base);
/* check repCode */
{ const U32 repIndex = (U32)(current+1 - offset_1);
const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
const BYTE* const repMatch = repBase + repIndex;
if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex)) /* intentional overflow */
if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
/* repcode detected we should take it */
const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
if (depth==0) goto _storeSequence;
} }
/* first search (depth 0) */
{ size_t offsetFound = 999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
2017-09-02 01:28:35 +00:00
if (ml2 > matchLength)
matchLength = ml2, start = ip, offset=offsetFound;
}
if (matchLength < 4) {
ip += ((ip-anchor) >> kSearchStrength) + 1; /* jump faster over incompressible sections */
2017-09-02 01:28:35 +00:00
continue;
}
/* let's try to find a better solution */
if (depth>=1)
while (ip<ilimit) {
ip ++;
current++;
/* check repCode */
if (offset) {
const U32 repIndex = (U32)(current - offset_1);
const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
const BYTE* const repMatch = repBase + repIndex;
if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex)) /* intentional overflow */
if (MEM_read32(ip) == MEM_read32(repMatch)) {
/* repcode detected */
const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
int const gain2 = (int)(repLength * 3);
int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
if ((repLength >= 4) && (gain2 > gain1))
matchLength = repLength, offset = 0, start = ip;
} }
/* search match, depth 1 */
{ size_t offset2=999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offset2);
2017-09-02 01:28:35 +00:00
int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
if ((ml2 >= 4) && (gain2 > gain1)) {
matchLength = ml2, offset = offset2, start = ip;
continue; /* search a better one */
} }
/* let's find an even better one */
if ((depth==2) && (ip<ilimit)) {
ip ++;
current++;
/* check repCode */
if (offset) {
const U32 repIndex = (U32)(current - offset_1);
const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
const BYTE* const repMatch = repBase + repIndex;
if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex)) /* intentional overflow */
if (MEM_read32(ip) == MEM_read32(repMatch)) {
/* repcode detected */
const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
int const gain2 = (int)(repLength * 4);
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
if ((repLength >= 4) && (gain2 > gain1))
matchLength = repLength, offset = 0, start = ip;
} }
/* search match, depth 2 */
{ size_t offset2=999999999;
size_t const ml2 = searchMax(ms, ip, iend, &offset2);
2017-09-02 01:28:35 +00:00
int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
if ((ml2 >= 4) && (gain2 > gain1)) {
matchLength = ml2, offset = offset2, start = ip;
continue;
} } }
break; /* nothing found : store previous solution */
}
/* catch up */
if (offset) {
U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */
offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
}
/* store sequence */
_storeSequence:
{ size_t const litLength = start - anchor;
ZSTD_storeSeq(seqStore, litLength, anchor, (U32)offset, matchLength-MINMATCH);
2017-09-02 01:28:35 +00:00
anchor = ip = start + matchLength;
}
/* check immediate repcode */
while (ip <= ilimit) {
const U32 repIndex = (U32)((ip-base) - offset_2);
const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
const BYTE* const repMatch = repBase + repIndex;
if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex)) /* intentional overflow */
if (MEM_read32(ip) == MEM_read32(repMatch)) {
/* repcode detected we should take it */
const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset history */
ZSTD_storeSeq(seqStore, 0, anchor, 0, matchLength-MINMATCH);
2017-09-02 01:28:35 +00:00
ip += matchLength;
anchor = ip;
continue; /* faster when present ... (?) */
}
break;
} }
/* Save reps for next block */
rep[0] = offset_1;
rep[1] = offset_2;
2017-09-02 01:28:35 +00:00
/* Return the last literals size */
return (size_t)(iend - anchor);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_greedy_extDict(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_lazy_extDict(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_lazy2_extDict(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
2017-09-02 01:28:35 +00:00
}
size_t ZSTD_compressBlock_btlazy2_extDict(
ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
void const* src, size_t srcSize)
2017-09-02 01:28:35 +00:00
{
return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
2017-09-02 01:28:35 +00:00
}